Detecting Drinks at a Crime Scene with Smart Cameras
Researchers harnessed hyperspectral cameras—capable of capturing 204 distinct color bands from visible to near‑infrared light—to investigate how different drinks leave fingerprints on surfaces. A staged crime scene yielded nine distinct beverage stains: papaya, coffee, pomegranate, orange, tea, wine, whisky, rum, and brandy.
Data Refinement
To distill the most informative wavelengths, the team applied an ANOVA statistical test, trimming the dataset from 204 to 162 bands and eliminating redundant noise that could skew analysis.
AI Models in Play
Four artificial‑intelligence architectures were trained on the refined spectra:
- Multilayer Perceptron (MLP) – a classic feed‑forward neural network.
- One‑Dimensional Convolutional Network (1D‑CNN) – captures local spectral patterns.
- Long Short‑Term Memory Network (LSTM) – models sequential dependencies across wavelengths.
- Hybrid CNN‑LSTM – combines convolutional feature extraction with temporal sequencing.
Training employed adaptive learning rates and early stopping to prevent overfitting. A five‑fold cross‑validation ensured each model was evaluated on varied data splits, yielding robust performance metrics.
Results
| Model | Accuracy |
|---|---|
| MLP | 95.58 % |
| 1D‑CNN | ~93 % |
| LSTM | ~92 % |
| CNN‑LSTM | ~93.5 % |
The MLP outperformed all others, achieving the highest stain‑recognition rate.
Implications
Combining hyperspectral imaging with deep learning offers a non‑destructive method to detect and classify drink stains, preserving crime scene integrity while extracting detailed forensic evidence. This technique holds promise for future investigative workflows.
